Research map / Continual learning / Continual fine-tuning of LLMs
Continual Model Routing in Evolving Model Hubs
Jack Bell, Giacomo Carfi', Gerlando Gramaglia, Vincenzo Lomonaco
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add embedding anchoring (ablation) | CARvE (10% domain replay) · CMRBench avg | M-Acc / D-Fgt: 41.7 -> 43.1 M-Acc; 7.4 -> 5.9 D-Fgt | Improved |
| add projection anchoring (ablation) | CARvE (10% domain replay) · CMRBench avg | M-Acc / D-Fgt: 40.1 -> 43.1 M-Acc; 9.4 -> 5.9 D-Fgt | Improved |
| use domain-stratified coreset replay instead of random replay at 5% budget (ablation) | CARvE · CMRBench avg | M-Acc: 37.3 -> 40.5 M-Acc; 77.4 -> 78.5 D-Acc | Improved |
| use domain coreset replay instead of random replay at 10-20% budget (ablation) | CARvE · CMRBench avg | D-Acc: 80.6 -> 80.7 D-Acc at 10%; 82.8 -> 82.9 at 20% | No clear effect |
| increase candidate set size K from 32 to 512 (ablation) | CARvE · CMRBench avg | D-Acc: 80.1 -> 81.0 D-Acc | No clear effect |
| corrupt 10% of training labels (ablation) | CARvE (10% domain replay) · CMRBench avg | D-Acc: 80.7 -> 72.1 D-Acc; 5.9 -> 8.0 D-Fgt | Hurt |
| corrupt 20% of training labels (ablation) | CARvE (10% domain replay) · CMRBench avg | D-Acc: 80.7 -> 63.1 D-Acc | Hurt |
| replace random replay with CARvE at 10% replay | LoRA LLaMA2-7B routing · CMRBench avg | D-Acc / D-Fgt: 75.9 -> 80.7 D-Acc; 13.1 -> 5.9 D-Fgt | Improved |
| initialise model embeddings from model-card representations instead of randomly (ablation) | CARvE · CMRBench | D-Acc: 3-5pp lower D-Acc, roughly double forgetting | Hurt |
| augment query with top-1 retrieved model card (RAT) instead of zero-shot (ablation) | fine-tuned LLaMA2-7B router · CMRBench | accuracy | Hurt |
| add EWC regularisation to CARvE (ablation) | CARvE (10% replay) · CMRBench avg | D-Acc: 80.7 -> 80.5 | No clear effect |
| use a smaller backbone (Qwen3-4B instead of LLaMA2-7B) (ablation) | CARvE (10% replay) · CMRBench avg | M-Acc: 43.1 -> 36.7 M-Acc | Hurt |
Closest papers by meaning
- CONCUR: A Framework for Continual Constrained and Unconstrained Routing (ICLR 2026 · similarity 0.88)
- Continual Model Merging without Data: Dual Projections for Balancing Stability and Plasticity (NeurIPS 2025 · similarity 0.85)
- Routing Experts: Learning to Route Dynamic Experts in Existing Multi-modal Large Language Models (ICLR 2025 · similarity 0.85)
- Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · similarity 0.84)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.83)
- Confidence-Guided Stepwise Model Routing for Cost-Efficient Reasoning (AAAI 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
